AI has made software generation dramatically more efficient. Increasingly, the constraint lies elsewhere.

Faros AI reports that epics completed per developer have increased 66.2% with AI. Over the same period, median pull request review time has increased 441%, while 31% more pull requests merge without review. Rather than creating a linear increase in delivery, faster code generation is exposing new constraints in how engineering organizations validate and integrate change.

The challenge is less about producing software than about establishing confidence that it belongs in the system.

AI changes the pace. Delivery determines the outcome.

AI adoption is now widespread. Pragmatic Engineer reports that 95% of engineers use AI tools weekly, while SecondTalent found that 72% of CIOs believe those investments are delivering little or no net return.

At the same time, Jellyfish found that top-quartile AI adopters achieve roughly twice the pull request throughput of lower-performing organizations using comparable tooling.

The difference is not the model. It is the operating model around it.

Organizations realizing sustained gains are redesigning how engineering work moves from request to production. Those seeing limited returns are often asking existing delivery processes to absorb significantly more generated code without changing how engineering decisions are made.

Review is becoming the new constraint.

Writing code has become substantially faster. Understanding that code has not.

Historically, the effort required to implement a change also created the context needed to evaluate it. Engineers developed an understanding of surrounding architecture, existing capabilities, and system assumptions as they worked.

AI shortens implementation time, but it does not shorten the reasoning required to determine whether a change is appropriate for the broader system.

That responsibility increasingly shifts to review.

As more code is produced in less time, reviewers spend proportionally more effort reconstructing context that was never established before implementation. The result is not simply longer reviews, but a growing mismatch between generation speed and validation capacity.

High-performing organizations move validation upstream.

The organizations capturing the greatest productivity gains are changing where engineering decisions occur.

Instead of relying on pull request review to determine whether a change should exist, they establish architectural context before implementation begins. Requests are evaluated against existing capabilities, technical constraints, and system patterns before an AI session generates code.

That changes the role of review.

Rather than serving as a discovery process, review becomes confirmation that an approved approach has been implemented correctly. Validation shifts from reconstructing intent after the fact to verifying execution against a shared understanding established earlier in the delivery lifecycle.

Consistency matters just as much as sequencing. Approval gates create value when every change passes through the same decision process, allowing reviewers to build confidence in what has already been evaluated instead of repeating that work independently.

AI rewards organizations that redesign delivery.

The most significant gains from AI are increasingly coming from changes to engineering systems rather than improvements in code generation itself.

McKinsey has similarly observed that organizations redesigning software delivery around continuous collaboration between engineers and AI systems are achieving productivity improvements of three to five times, substantially outperforming organizations that layer AI onto existing development practices.

As AI continues to reduce the cost of producing software, engineering performance will increasingly depend on how efficiently organizations establish shared context, make architectural decisions, and validate change—not simply how quickly code can be generated.

‍